[llvm] 866b195 - [AMDGPU] Regenerate checks for waitcnt-overflow.mir

Jay Foad via llvm-commits llvm-commits at lists.llvm.org
Fri Dec 31 03:31:40 PST 2021


Author: Jay Foad
Date: 2021-12-31T11:27:15Z
New Revision: 866b195cb9d7205de25d5df1eeb1cbf7229619f4

URL: https://github.com/llvm/llvm-project/commit/866b195cb9d7205de25d5df1eeb1cbf7229619f4
DIFF: https://github.com/llvm/llvm-project/commit/866b195cb9d7205de25d5df1eeb1cbf7229619f4.diff

LOG: [AMDGPU] Regenerate checks for waitcnt-overflow.mir

Added: 
    

Modified: 
    llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir

Removed: 
    


################################################################################
diff  --git a/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir b/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir
index eb8dc27759899..3b48d23cd38b0 100644
--- a/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir
+++ b/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir
@@ -21,60 +21,60 @@ body:             |
 
     ; GFX9-LABEL: name: max-counter-lgkmcnt
     ; GFX9: S_WAITCNT 0
-    ; GFX9: $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
-    ; GFX9: $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
-    ; GFX9: $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
-    ; GFX9: $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
-    ; GFX9: $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
-    ; GFX9: $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
-    ; GFX9: $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
-    ; GFX9: $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
-    ; GFX9: $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
-    ; GFX9: $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
-    ; GFX9: $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
-    ; GFX9: $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
-    ; GFX9: $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
-    ; GFX9: $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
-    ; GFX9: $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
-    ; GFX9: $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
-    ; GFX9: $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
-    ; GFX9: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
-    ; GFX9: S_WAITCNT 52863
-    ; GFX9: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
-    ; GFX9: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
-    ; GFX9: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
-    ; GFX9: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
-    ; GFX9: S_ENDPGM 0
+    ; GFX9-NEXT: $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
+    ; GFX9-NEXT: S_WAITCNT 52863
+    ; GFX9-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX9-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
+    ; GFX9-NEXT: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
+    ; GFX9-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
+    ; GFX9-NEXT: S_ENDPGM 0
     ; GFX10-LABEL: name: max-counter-lgkmcnt
     ; GFX10: S_WAITCNT 0
-    ; GFX10: S_WAITCNT_VSCNT undef $sgpr_null, 0
-    ; GFX10: $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
-    ; GFX10: $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
-    ; GFX10: $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
-    ; GFX10: $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
-    ; GFX10: $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
-    ; GFX10: $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
-    ; GFX10: $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
-    ; GFX10: $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
-    ; GFX10: $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
-    ; GFX10: $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
-    ; GFX10: $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
-    ; GFX10: $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
-    ; GFX10: $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
-    ; GFX10: $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
-    ; GFX10: $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
-    ; GFX10: $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
-    ; GFX10: $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
-    ; GFX10: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
-    ; GFX10: S_WAITCNT 53631
-    ; GFX10: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
-    ; GFX10: S_WAITCNT 53375
-    ; GFX10: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
-    ; GFX10: S_WAITCNT 53119
-    ; GFX10: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
-    ; GFX10: S_WAITCNT 52863
-    ; GFX10: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
-    ; GFX10: S_ENDPGM 0
+    ; GFX10-NEXT: S_WAITCNT_VSCNT undef $sgpr_null, 0
+    ; GFX10-NEXT: $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
+    ; GFX10-NEXT: S_WAITCNT 53631
+    ; GFX10-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX10-NEXT: S_WAITCNT 53375
+    ; GFX10-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
+    ; GFX10-NEXT: S_WAITCNT 53119
+    ; GFX10-NEXT: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
+    ; GFX10-NEXT: S_WAITCNT 52863
+    ; GFX10-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
+    ; GFX10-NEXT: S_ENDPGM 0
     $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
     $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
     $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
@@ -109,155 +109,155 @@ body:             |
 
     ; GFX9-LABEL: name: max-counter-vmcnt
     ; GFX9: S_WAITCNT 0
-    ; GFX9: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
-    ; GFX9: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
-    ; GFX9: S_WAITCNT 53118
-    ; GFX9: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
-    ; GFX9: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
-    ; GFX9: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
-    ; GFX9: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
-    ; GFX9: S_ENDPGM 0
+    ; GFX9-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
+    ; GFX9-NEXT: S_WAITCNT 53118
+    ; GFX9-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX9-NEXT: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
+    ; GFX9-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
+    ; GFX9-NEXT: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
+    ; GFX9-NEXT: S_ENDPGM 0
     ; GFX10-LABEL: name: max-counter-vmcnt
     ; GFX10: S_WAITCNT 0
-    ; GFX10: S_WAITCNT_VSCNT undef $sgpr_null, 0
-    ; GFX10: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
-    ; GFX10: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
-    ; GFX10: S_WAITCNT 65406
-    ; GFX10: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
-    ; GFX10: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
-    ; GFX10: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
-    ; GFX10: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
-    ; GFX10: S_ENDPGM 0
+    ; GFX10-NEXT: S_WAITCNT_VSCNT undef $sgpr_null, 0
+    ; GFX10-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
+    ; GFX10-NEXT: S_WAITCNT 65406
+    ; GFX10-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX10-NEXT: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
+    ; GFX10-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
+    ; GFX10-NEXT: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
+    ; GFX10-NEXT: S_ENDPGM 0
     $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
     $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
     $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
@@ -341,29 +341,29 @@ body:             |
 
     ; GFX9-LABEL: name: max-counter-expcnt
     ; GFX9: S_WAITCNT 0
-    ; GFX9: EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX9: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
-    ; GFX9: S_ENDPGM 0
+    ; GFX9-NEXT: EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX9-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX9-NEXT: S_ENDPGM 0
     ; GFX10-LABEL: name: max-counter-expcnt
     ; GFX10: S_WAITCNT 0
-    ; GFX10: S_WAITCNT_VSCNT undef $sgpr_null, 0
-    ; GFX10: EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
-    ; GFX10: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
-    ; GFX10: S_ENDPGM 0
+    ; GFX10-NEXT: S_WAITCNT_VSCNT undef $sgpr_null, 0
+    ; GFX10-NEXT: EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
+    ; GFX10-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX10-NEXT: S_ENDPGM 0
     EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
     EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
     EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec


        


More information about the llvm-commits mailing list